Govcaster

Web3 governance suffers from extreme voter apathy. We aim to address this issue by creating governance-related frames for Warpcast. Using generative AI and a finetuned LLM, we create attention-grabbing headlines and thought-provoking questions, as well as customized assets to promote governance interest and activity on Warpcast. Govcaster uses data science methods to identify relevant and influential voices on governance fora and succintly distribute their posts and activity to Warpcast users. Utilizing the innovations in frames, we have also embedded the ability for users to delegate, vote, and learn more about a governance proposal, all without ever leaving the platform.

We began with the backend, which is powered by the Discourse API and OpenAI. We acquire raw forum data for Optimism using Discourse, and then perform a series of analyses to continuously identify “high-signal” users. We track these users and their activity and use statistical methods to determine at what times to make a post related to their activity. We also identify “high-relevance” posts, which may or may not be made by high-signal users. When either a post or a series of user activity is identified as publication-worthy, we make a call to GPT-3.5 to summarize, inquire about, or directly quote the input in such a way that it is compatible with a frame. (Note that we are currently developing a finetuned instance of GPT-4 for when we ship this product. We are also experimenting with DALLE-3 and StableDiffusion as two options for custom assets in the frames when we ship.)

On the frontend, we store all assets in Pinata. For the purpose of this hackathon, the assets are static, but as mentioned above, in production, they will be dynamic based on the nature and sentiment of the proposal or forum activity. We used Figma for mockups and converted whatever we had time for into next.js code that we deployed to Vercel. The remainder was uploaded as screenshots to Pinata and rendered as such for demo purposes.